Feng Wang

Misrouter introduces an input-only adversarial framework to exploit the routing mechanisms of Mixture-of-Experts (MoE) LLMs, enabling unsafe behavior induction against remotely hosted, black-box services.

The paper introduces SCAgent, an automated framework that uses LLM-assisted agents to systematically discover, analyze, and assess side-channel leakage risks in complex systems like iOS, moving beyond manual and predefined analysis.

The paper proposes BiRD, a bidirectional ranking defense mechanism that enhances the robustness of Retrieval-Augmented Generation (RAG) against adversarial attacks by analyzing the alignment between forward and backward document rankings.

The paper introduces MOV-Bench, a challenging benchmark for multi-hop audio-visual reasoning, and proposes AOP-Agent, an agentic framework that significantly improves open-source Omni-LLMs' ability to perform active cross-modal perception.

C-MIG is a novel retrieval-augmented generation framework that uses multi-view information gain to improve clinical diagnosis reasoning by providing richer, more nuanced reward signals than existing methods.

The paper proposes EAPO, an entropy-driven adaptive weighting method that dynamically adjusts the influence of positive samples during policy optimization to improve both response diversity and stability in open-ended QA.

The paper introduces AgentSchool, an advanced LLM-powered multi-agent simulator that models learning as state transitions to provide a robust, ethically viable testbed for educational research and pedagogical reform.

OptSkills introduces an archetype-centric skill learning agent that improves the generalization of solving optimization problems from natural language by clustering problems by underlying archetypes and distilling reusable workflow skills.

The paper introduces Singularity-aware Adam (S-Adam), a novel optimizer that stabilizes deep learning training in non-smooth loss landscapes by dynamically damping updates based on local geometric instability.

The paper introduces SkillBrew, a multi-objective framework that treats skill bank curation as a constrained optimization problem to build efficient and well-curated skill repositories for LLM agents.

The paper proposes EKSFT, a selective fine-tuning method that masks high-entropy or high-KL divergence tokens during Supervised Fine-Tuning (SFT) to prevent distribution shift and improve subsequent Reinforcement Learning (RL) performance.

UniAudio-Token is a framework that enhances existing semantic speech tokenizers with general audio perception, allowing them to handle diverse audio types while maintaining high-fidelity speech capabilities.

This paper investigates the 'faithfulness gap' in LLM agents—the discrepancy between stated reasoning and actual action—by decomposing it into two opposing steps: reasoning-to-conclusion and conclusion-to-action.

The paper proposes LaSR, a context-aware training paradigm that uses latent reasoning to significantly improve speech recognition, especially for specialized terminology, without adding latency.

The paper introduces DiscourseFlip, a novel graph-guided attack that demonstrates how coordinated poisoning across a multi-topic query space can manipulate the overall opinion generated by black-box Retrieval-Augmented Generation (RAG) systems.

The paper demonstrates that enforcing a local conservative finite volume structure is crucial for achieving stable, accurate long-term autoregressive rollouts of plasma transport simulations, outperforming learned neural network approaches.

The paper introduces DiscourseFlip, a novel black-box, graph-guided attack that manipulates opinions across an entire multi-topic query network, demonstrating a significant leap in scope and effectiveness over existing RAG attack methods.

The paper introduces MCP-Persona, a novel benchmark designed to evaluate LLM agents' performance on real-world, personalized applications using the Model Context Protocol (MCP), revealing that current state-of-the-art agents struggle with such personalized tool use.

dstack-capsule is a Kubernetes platform that enables fine-grained, Pod-level remote attestation on Intel TDX, allowing multiple confidential workloads to share a single VM without sacrificing security or incurring excessive resource overhead.

This paper introduces MedReCo and MedReCo-VLM, a framework that enables entity-aware cross-image reasoning for medical imaging, allowing AI to compare current scans with prior studies and analogous cases based on structured clinical reports.